mirror of
https://github.com/wassname/ml_debug.git
synced 2026-09-06 16:50:17 +08:00
wassnameandClaudypoo
3a58c54170
- refs/ -> references/, the folder name the Agent Skills spec uses and the one Hermes skips when it walks for nested skills. - rl and pinn declared name: rl and name: pinn, which are global names in a flat skill namespace. Now ml-debug-rl and ml-debug-pinn. They also called themselves sub-skills of 'ml-debugging', which is not this skill's name. - Drop the dead link to SKILL_old.md. It moved into gitignored slop/, so the link was broken for anyone who cloned. - Route references/llm_judge_litreview.md, the one reference SKILL.md never named. - Description leads with the trigger situations. Hermes truncates it to 57 chars. Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
1.1 KiB
1.1 KiB
Local Instructions
Quote-first evidence
When adding evidence or appendix material for this skill, prefer the expert's own words over assistant synthesis. These sources are high-level SME material, often above current frontier LLM taste and more diverse than model priors. Quoting them is grounding data; rewording them injects assistant bias.
- Use generous quote blocks, usually 2-3 sentences or a full paragraph.
- Preserve the author's paragraph flow. Do not insert blank blockquote lines between sentences from the same paragraph; it makes expert prose harder to read and adds fake structure.
- Keep editorial text sparse and mostly for routing: why this source is here, when to read it, and what local file it supports.
- Do not atomize useful source material into tiny one-line quotes when a longer block preserves the author's reasoning.
- Put lower-relevance sources in "See also" rather than forcing a synthetic narrative around them.
- In
SKILL.md, link to reference docs likereferences/research_taste.mdinstead of copying a long assistant-written summary.